Google has quietly killed Project Mariner, the agentic browser tool that could fill out forms, book travel and click through web tasks on a user's behalf, and is folding its capabilities into the Gemini assistant. There was no headline announcement, but the move lands just before Google I/O and signals a sharp rethink of how the company wants its agents to operate.
Mariner was Google's answer to products like Anthropic's Claude coworker: a model that visually parses a browser, takes screenshots and drives a cursor through the document object model. On paper it was a flagship demo of the agent era. In practice, the unit economics did not hold up. "What I think is actually happening here is that Mariner basically didn't make unit economics sense. The browser agent that can take, you know, it's taking screenshots and it's visually parsing the DOM. They're very slow. They're very expensive," Jaeden Schafer said on the podcast.
The competitive context matters. Command-line agents such as OpenAI's Codex and Anthropic's Claude Code have converged on much of the same functionality without the overhead of pixel-level vision, and Google appears to have read the same data. Schafer argued the company is watching how rival agents are architected and adjusting accordingly rather than continuing to subsidise an expensive vision-first stack.
Key facts
- 01Google has shut down Project Mariner, its agentic browser tool, and is folding the underlying tech into the Gemini agent ecosystem ahead of Google I/O.
- 02A recent report cited on the show pegs computer-use agents at roughly 85 times as compute-intensive as text-based coding agents like Claude Code or Codex.
- 03Mariner overlapped heavily with Anthropic's Claude coworker, which can take screenshots, parse the DOM and click through apps such as Audacity.
- 04Code-first agents increasingly bypass the GUI entirely — editing a CapCut project file directly rather than driving the CapCut app via screenshots.
The cost gap is not subtle. There was a recent report that came out and it said that doing that, which is called computer use is about 85 times as compute intensive as doing something through just directly through Claude Code or, you know, asking it questions and it kind of giving you a text response, Jaeden Schafer said. At that ratio, every screenshot-driven session is multiples more expensive to serve than a comparable text or code task — a hard problem when agents are meant to run continuously in the background.
“Mariner basically didn't make unit economics sense. The browser agent that can take, you know, it's taking screenshots and it's visually parsing the DOM. They're very slow. They're very expensive.”— Jaeden Schafer
Computer-use agents do unlock genuinely useful workflows that text agents cannot match. Schafer described handing Claude coworker control of his screen to run chains of edits inside Audacity, the open-source audio editor, with the model clicking through menus and adjusting settings as a human would. The capability is real; the question is whether it can be priced competitively against narrower alternatives.
The emerging workaround is to skip the GUI altogether. Schafer noted that Claude Code and Claude coworker increasingly try to solve tasks by manipulating files directly rather than driving applications. Asked to edit a video in CapCut, the agent has had more success opening the underlying project file and rewriting it in code than launching the CapCut app and clicking through the timeline. It is, as he put it, a "crazy robot way to do things," but it sidesteps the screenshot tax entirely.
For Google, folding Mariner into Gemini concentrates agent work inside a product with a billion-plus user funnel and a clearer monetisation path, while quietly retiring a brand that had become synonymous with the slow, screenshot-driven approach. It also gives the Gemini team room to mix vision-based control with cheaper code-first execution depending on the task, rather than committing to one paradigm.
The broader signal is that the first wave of browser-piloting agents is getting repriced against reality. Demos that wow on stage have to clear an inference bill at scale, and an 85x compute multiple is hard to absorb when rivals can deliver similar outcomes through code. Mariner's quiet shutdown suggests even the company that helped popularise the category is not willing to keep paying that premium.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
